Recognizing Textual Entailment by Soft Dependency Tree Matching

Abstract

We present a rule-based method for recognizing entailment relation between a pair of text fragments by comparing their dependency tree structures. We used a dependency parser to generate the dependency triples of the text–hypothesis pairs. A dependency triple is an arc in the dependency parse tree. Each triple in the hypothesis is checked against all the triples in the text to find a matching pair. We have developed a number of matching rules after a detailed analysis of the PETE dataset, which we used for the experiments. A successful match satisfying any of these rules assigns a matching score of 1 to the child node of that particular arc in the hypothesis dependency tree. Then the dependency parse tree is traversed in post-order way to obtain the final entailment score at the root node. The scores of the leaf nodes are propagated from the bottom of the tree to the non-leaf nodes, up to the root node. The entailment score of the root node is compared against a predefined threshold value to make the entailment decision. Experimental results on the PETE dataset show an accuracy of 87.69% on the development set and 73.75% on the test set, which outperforms the state-of-the-art results reported on this dataset so far. We did not use any other NLP tools or knowledge sources, to emphasize the role of dependency parsing in recognizing textual entailment.